Prith Banerjee: Synopsys's Bold Vision for the Future of Engineering

Great to be on this podcast with you. Great. I'm looking forward to it.

Yes. So Synopsys is a leading provider of electronic design automation tools, IP, and simulation. When you have a chip from either Nvidia, Intel, Qualcomm, Broadcom, those chips are designed using electronic design automation tools from Synopsys. And then these large chips have significant intellectual property blocks in them. Blocks like PCIe Express or Ethernet controllers or memory controllers. Those are also provided by Synopsys. And then with the latest acquisition of Ansys, a leading simulation provider, we now have the full chip-to-systems solution from Synopsys.

Actually, that's an interesting question. You mentioned my bio. I used to be a professor, so I'll actually start at the beginning. I started as a professor at the University of Illinois, Urbana-Champaign, teaching VLSI design. This is how do you actually teach students how to design VLSI chips using Carver Mead's book on VLSI design. At that point, it was all custom IC design. I taught students to actually draw rectangles on a screen. Those rectangles represent masks. And so you have a polysilicon mask over a diffusion mask that implements a transistor. That's how chips used to be designed in the 80s. But custom IC design had its limitations. You could only design maybe a human designer could do design maybe 10,000 rectangles, 100,000 rectangles. So about 10,000 transistors. That industry took a big move through the use of what is called standard cell design, where you pre-design cells for a two-input NAND gate, four-input NAND gate, etc., small, medium, large versions. And so instead of drawing rectangles, designers would draw gate-level netlists, right? AND gate, connected NOR, and so on. And then when I started teaching VLSI design, I was teaching it at that higher level. And so the electronic design automation tools, the EDA industry moved towards placement and routing of those cells on a chip. Then came the next big deal, actually by Synopsys, which is RTL synthesis, where you enter a design at a register transfer level, Verilog or VHDL, and then the Design Compiler, this synthesis tool, generated the netlist of gates, and then you did the placement routing to do the rectangles. So something that a human designer could do with 100,000 transistors, now you could do a million transistors with placement and routing, now with 10 million transistors, and so on. So this whole industry was significantly impacted with the invention of RTL synthesis, and that was the beginning of Synopsys, and I was part of the journey. I used to be a professor coming to the ICCAD and DAC conferences, and I met Art, the founder, and Alberto Sangiovanni-Vincentelli. So these are two of the founders that I have known for many, many years. And so this has been an incredible journey. But then the whole industry said, okay, I want to do even more complex things, right? So essentially you added more IP blocks, pre-designed IP blocks. So instead of doing standard cell gates, you essentially took a memory IP block or an ARM processor IP block or an Ethernet controller IP block. So in a large chip with 100 billion transistors, you may have 10, 15 IP blocks, and only a small chip is done through electronic design. So that's essentially what was happening in the chip industry. But then as the amount of transistors that you could put on a chip got sort of maxed out, right? Because the size of your chip is fixed. So you can only put that much in it, right? So essentially people said, hey, how can I put more transistors in a chip? So you say take a small chip, which is a chiplet, and then you put multiple chiplets together on a die, and you connect them up through a 3D IC and so on. So now that is what is going on. You are jamming more and more transistors inside a multi-chip thing with 3D ICs. So the complexity of that is getting even more complicated, and that's where the Ansys part comes in, because with Ansys you can do all the multiphysics simulation for the thermals and electromagnetics and so on, and we'll talk about it. But it is an incredible journey, and now people are talking about putting a trillion transistors on a cheaper wafer. The Cerebras chip has, I think, 2.6 trillion transistors on it. So it's an incredible journey going from the 100,000 rectangles that I used to teach in 1984 to the trillion transistors today.

Yes. So the rationale behind the deal is the fact that the intelligent systems that we see around us, right? In any industry, from either automotive, aerospace, healthcare, energy, industrials, every one of those industries has... So suppose you're looking at a car, that car is a software-defined car, right? So there's a piece of some 10 million lines of code that is running on that car, on that chip to make that car autonomous or semi-autonomous. That 10 million lines of code or 100 million lines of code runs on an ECU. It's a silicon chip, right? Designed using EDA tools from Synopsys. But then that chip, that software is doing the autonomous control of a vehicle. That vehicle is an electromechanical part, right? It has wheels in it. So with Ansys simulation, you can actually model the inertia, the mechanics of it, the external aerodynamics, how much wind drag there will be, etc., right? And then if it crashes into a wall, what will be the crashing? So it's an electromechanical thing that is going on that is being controlled by the software running on a chip. So in every industry, the vision is it's software-defined, AI-enabled silicon chip design, right? This is the transformation. And so to make it successful, you need to inside the same design tools. What if you could get the chip design as well as system design in the whole chip-to-systems? That was the thesis behind it, right? So earlier Synopsys was in the area of electronic design automation. What's the size of the market? It's $10 billion, $15 billion, right? Now the market is the entire product area, right? So collectively across industries, the market is like $1.3 trillion. That's the R&D budget. That has essentially become the market for these tools to design products in this space. That was the rationale. So then the transaction was announced in January of 2024. It took 18 months for the transaction to go through with all the regulatory approvals and so on. And we were all excited about it on both sides, and lots of integration discussions were happening, but we actually couldn't really do it because until the thing is done. I have now been part of Synopsys, the executive leadership team, and from July we are now the combined company. There's some very, very interesting joint solutions that are being built together, right? So for example, if you have the sort of this multi-chip, the 3D IC that I was talking about, right? You put multiple chiplets on a chip, and essentially you're trying to find out what the thermal behavior will be, what the EMI, the electromagnetic will be, right? We have electromagnetic tools from Ansys HFSS that can do really detailed electromagnetic interactions, or we had power tools from Ansys which is RedHawk. So that RedHawk can be integrated with the EDA tools PrimeTime and so on, right? To do this. So we have done some absolutely phenomenal joint solutions work that is going on, and our customers are loving those joint solutions.

It's a great question, actually. So if I need to kind of explain to you why this kind of computational power is needed, right? So when you look at an electronic design automation tool like say Fusion Compiler that automatically sort of synthesizes the design, right, from RTL all the way to placement and routing, or PrimeTime which does the sort of timing sign-off. These computations require a lot of time to run, right? Or then you look at the Ansys simulations, right? Fluid simulation, I mean, this is partial differential equations. You have to solve them numerically. It can take a thousand hours to run a fluid simulation or a mechanical simulation doing crash things. And so these things require a lot of compute. And so in these areas, right, when you try to run them fast, right, on a well, when your clock frequency of your chip was increasing every 18 months, you are getting double the frequency. You could just wait for the next processor to come in and it would run faster. But as the frequency has sort of plateaued at 2 GHz, you have to rely on something else to make it happen. So the early work was, well, let's just have multiple cores within a CPU and you rely on instruction-level parallelism to dynamic scheduling to get speedups. That kind of took us to maybe two times, three times the speedup. Then you had said, hey, now I have to rewrite my code to take advantage of shared memory parallelism. Meaning I take my code, I say fork a process here, fork another process here, and then you join the processes and then you lock, unlock. So you're writing programs to do shared memory programming. That took you to maybe 15 times, 16 times speedups. It was not scaling to hundreds. So next step was do message passing, distributed memory. The reason I'm going to it, this is my research background. This is what I used to work on as an academic, as a professor. So the message passing part let us go up to multiple things, right? Maybe hundreds of processors. So both the Synopsys part and the Ansys part of the tools, we're working on that journey. But these codes, you could literally take these codes, taking an original Fortran code and say, well, I will make a shared memory version or a message passing version. But then the GPUs came around, right? And people said, wow, there's an interesting thing. GPUs first came around for just to do ray tracing, graphics, and so on. But both companies realized these things can be used to do parallel processing, but at a very, very fine grain, like 16,000 cores. Well, you cannot take a normal sequential algorithm with a Fortran program and say, oh, I want to run it on 16,000 cores. You have to really rethink how will Fluent be executed using GPUs, right? And a single GPU, multi-GPU. So we at Ansys, we are on a journey to take each of our codes, Fluent, HFSS, Mechanical, LS-DYNA, and get them running on GPUs. Same thing at Synopsys, right? Fusion Compiler, PrimeTime, TestMAX, right? So we are on a journey to make this go faster with GPUs. So the partnership with Nvidia is first, let's get this industrial software right with Synopsys, which Synopsys, Ansys, right? Each of these codes running efficiently on GPUs, which will completely change the way, right? You can do things much, much faster. But then there's this other opportunity of AI. Each of the tools, it turns out, be like for physics simulation, you can use AI to train them. And then Fluent powered with AI will run like a thousand times faster. So it's not GPU parallelism, it's just, well, you train the model using our platform like SimAI or OptiSlang, it runs so much faster. So what Nvidia noticed is, hey, the world is now being transformed into what is sort of physical AI, which is the world around us is autonomous driving or robotics and so on. And as the robot interacts with its environment, the environment interacts with is not words as tokens as ChatGPT or images as tokens as DALL-E or videos as tokens as in Sora. It's a whole new world. It's world physics models, right? And the physics models, you have to create the synthetic data. Who creates it? And Synopsys. So it made a lot of sense for Nvidia to take a look at this combined company, say, this is the company that Nvidia should be partnered with to take that industrial software, make it run first of all on GPUs, secondly run it with the AI-powered stuff, and then use that thing to train the next set of autonomous driving vehicles and the robots. So it's like the full circle. It's an incredible partnership, and our teams are working so fast, so hard on these joint solutions.

That is a great question. And again, let me share with your audience, right, the complexity that is changing, right? So there was a time when you had maybe a million transistors on a chip, then it became a hundred million, then a billion, now it's a hundred billion going to a trillion transistors on a chip, on the chip itself. But now look at the intelligent product that chip is running a 10 million line software code, and that software code is running on this chip, and that is controlling all these airplane wings and so on. So you look at any product in any industry, a car used to be designed, it took like six years or four years to design a car, and that you now need to design a car in two years or a one year. A chip that took maybe three years to design, now you need, I mean, Jensen asked for Nvidia to make a chip every year. So that type, the time to innovation, right, the next product is decreasing, and the complexity of the chip or the system has gone from 1 million to 10 million to 100 million to a billion, right? So the complexity is increasing, and the time is getting squished. So if a human engineer could say, do, I mean, just think about laying bricks on a wall, right? A human can lay a thousand bricks in a given year, right? So now that same engineer has to do instead of a thousand bricks, 10 billion bricks. So you need like thousands, millions of engineers to do the job. That is not practical. So what we have done is to say, hey, can AI help? So basically all the stuff that a human engineer was doing is, can I do it with AI? So the first step was we took each of our EDA tools and Ansys tools and we put AI inside those tools, right, with what we call machine learning enabled, right. So you took Fluent powered by AI or PrimeTime powered by AI. Then we said, let's take the whole stack of optimization. When you're running a chip design or whatever, you are actually doing various optimization things, and you're saying, if I try this, this is going to happen. If I try that, that is going to happen. So it turns out this is the whole world of what is called reinforcement learning. And you design. So the early work was on design space exploration. When you're doing digital design, you're doing this, turning this 15 different knobs in all these different directions. And just like when you teach a child, you're learning how to ride a bike, you go left and you fall down, that's a penalty. You ride well, that's a reward. So you give a reward and a penalty. Use the reinforcement learning, and that is the core of how we did DSO. Once that worked for digital design, said, okay, can it work for verification? That was VSO. Then we said, does it work for test? That was TSO. So essentially we took the entire set of suite of EDA tools, DSO, VSO, ESO, etc., and we made all those things. The next step in the journey was, oh, the human designer, Venit, you're doing this, you are actually, you need to create a test bench. Can I have an assistive agent that will help me in creating the test bench? Can I create the test benches? So essentially now these agents are assistive agents or creative agents, and we have made a lot of things in the journey. The next step, which is the future, is, hey, all these things agents will work among themselves in a multi-agentic workflow and do the job of an entire team of engineers, full automation. So just like in an autonomous driving, you have level one driving, which is you're assisting the driver to do cruise control, right? Or level two is sort of automatic parking, right? And level five is fully autonomous driving, right? With these Waymo cars in San Francisco, we are imagining a world where the complete process of a chip design of this complexity of these trillion transistor chips in a year, less than a year, will be done with this agentic engineers. So we think we are excited. We have just talked about it in our Converge conference about re-engineering engineering. So how the process of engineering these complex chips to systems will be re-engineered with these agent engineers working hand in hand with human engineers.

This is actually a very, very interesting thing. I mean, I'll give you an example within the sort of 3D CAD computer area design area, right? So a human engineer working for a car company, right, would say, here is the design of the car, right? They will actually design the car shape on a CAD tool. And then they would use the Ansys simulation to simulate how the external fluid flow dynamics happens, or they will bang the car virtually, right, to see what happens, right? And if you don't like the design, so you did the CAD, the human designer did the CAD, then the simulation evaluated what that CAD does, right? And if you don't like it, the human designer went in and changed the CAD. Oh, now let's remove the mirror from here so there's less drag, right? And then you do a simulation. Oh, I liked it. It turns out AI is now saying, hey, you did the CAD this way. Let me do a CAD in a slightly different way. So we have a tool called GeoAI that makes suggestions to the changes in the geometry based on the objectives that you have set up. So there are literally thousands of different geometries that are being generated by GeoAI that are being simulated very fast with SimAI. So it's this GeoAI generates the geometries and SimAI simulates the things, and it's this combination of multi-agents that is generating these really, really unique things. Sometimes this geometric AI creates designs that a human designer has not thought of because a human designer gets tired. They only work eight hours a day. This thing works 24/7 non-stop, right? And it's like playing chess, right? When you're playing chess, a human can only think about two moves down. A computer can go like eight moves down, right? It has tremendous resources to explore. That is the power of AI.